0

I've got 3D voxel data, and I want to re-package it for memory efficiency and fast access. The data is generated in a regular octree, one integer value per cell. Unfortunately the data is not sparse, but the cells with the same value should be connected. 

Example for one slice:
[11122]
[11223]
[12222]
[44444]

My current idea is to use a kD-Tree, preferably left-balanced, but I'm not sure if there is an efficient algorithm to generate this. I've got some ideas, but I was hoping that this is one of those problems that already has established algorithms, or at least a name I could google for.

user1387
  • 187
  • 10

2 Answers2

1

How about OctoMap? As I understand, it's like an Octree, but merges adjacent occupied areas into regions to save memory. But I don't know much about it.

EDIT

You could also try my PH-Tree. It works like a octree, but is quite memory efficient because every node only stores bits that are different from the parent node. You could actually store your integer value as a 4th dimension. Contrary to intuition, a 4D tree may require less space than a 3D tree and it may be faster (explanation is in the PDF that can be found in the link above). If your integer is the 4th dimension, than any sub-tree in the tree will only have entries with 'similar' integers, may be that is sufficient for your case? Also, any node contains only close neighbours, but close neighbours are not necessarily in the same (or adjacent) nodes.

TilmannZ
  • 1,784
  • 11
  • 18
0

One further link: http://www.openvdb.org/ . Why did I only find this after asking the question? It's like asking for something in the supermarket only to find out that you're standing next to it.

I ended up doing something simpler, because I needed a solution: I convert the voxel volume into a stack of 2D planes, and each plane stores at which point the value changes to the next higher plane. That way the voxel data is only compacted vertically, but it seems to be "good enough" for now. I'll crunch the numbers (space requirement vs. performance) for other data structures if I have some free time.

user1387
  • 187
  • 10